计算机学报 杂志社,计算机学报

Abstract &

BackgroundAbstract Basing on a new scheme—random evolution plus feedback, which is reported to well represent the nature of biological evolution process, this paper proposes chaotic parallel genetic algorithm with feedback mechanism. In this new algorithm, chaotic mapping is embedded for maintaining a good diversity of population; and Baldwin effect based posterior reinforcement learning, which can successfully deal with the feedback information from the evolutionary system, is integrated to speed up the evolution along the right direction. The performance of this new algorithm was demonstrated on a well-known benchmark constrained non-linear problem and a benchmark problem of parameter estimation in finance. Experimental results and comparisons show that this new genetic algorithm is effective, universal and robust.

keywords genetic algorithm; chaos mapping; non-linear programming; Perato-dominant; Baldwin effect

background It is widely reported that standard genetic algorithm (SGA) confronts such problems as slow convergence and prematurity (crowding). Difficulty lies in that it is state-of-the-art to well balance the population diversity and selective pressure simultaneously. Lots of researches attempt to deal with these problems by enhancing the structure of individual or by designing complicated strategies for evolutionary operators. However, conventional improvements did not nearly solve these problems. Several reasons can explain it. Firstly, diversities of an initial population are far from actualized. Secondly, conventional initialization methods such as uniform and random approach can bring problems, too. Even if they can guarantee that the initial population is evenly distributed in search space, they cannot guarantee the qualities of initial population are also uniformly arranged. Thirdly, the diversity of population cannot be maintained under selective pressure. Finally and perhaps the most importantly, conventional GAs and their improvements completely abandon the individuals’ experiences during their lifetime, which causes no necessary connections between the current and next generations except for some controlling parameters. However, many experiments show that an improved GA with resource to domain-specific heuristics information always has a good performance in evolution. In the real biosystems, those who can successfully deal with the feedback information from evolution always survive better and keep evolving from low to higher classes.

This paper contributes to solving the problems of slow convergence and prematurity for GAs. The authors take advantages of chaotic system to enhance the diversity of population during evolution process and at the same time embed the reinforcement learning basing on heuristics information from evolutionary system into GA to speed up the convergence along the right direction.

Recently, they have published several papers try to solve the dilemma of GA. Paper titled "Inconsistent self-adaptive genetic algorithm and its application" appears in Systems Engineering(2002), introduced an inconsistent self-adaptive strategy into the operators of GA. Another paper titled "Baldwin effect based self-adaptive generalized genetic algorithm and its application" which appeared in the Proceedings of The Eighth International Conference on Control Automation, Robotics Vision (2004), proposed the Baldwin effect based reinforcement learning operator and introduced it into GA. Paper titled "Chaotic parallel genetic algorithm with feedback mechanism and its application in complex constrained problem" in Proceedings of IEEE Conference on Cybernetics and Intelligent Systems (2004), tried to combine both chaotic mapping and Baldwin effect based reinforcement learning operator into GA. However, the effect was not ideal as expected. In this paper, basing on the authors’ previous works they combine advantages of the inconsistent self-adaptive strategy, chaos system and the posterior reinforcement learning into genetic algorithm to build a novel Chaotic Parallel Genetic Algorithm with Feedback (CPGAF). Numerical experimental results show that the new genetic algorithm is effective, universal and robust.